RB Compound · agents for Claude Code

RB Compound: Enterprise grade agents suite

six agents for Claude Code that write before they code: each feature gets a story and a design, you approve the design, and only then do the agents write the code, tests first. every step leaves a Markdown file in your repository.

license PolyForm Noncommercial 1.0.0: free for noncommercial use.

  1. the epicepicman
  2. the storiesstoryman
  3. the questionsbsman
  4. the designdesignman
  5. you approveand only then, the code:
    1. tddman
  6. the docsdocman
// RB Compound, the agents

six agents, from the epic to the docs.

RB Compound is a suite of agents for Claude Code. each one does one job and leaves a written file behind: the next agent starts from that file, and so do you.

1
epicman
cuts the project into stories
2
storyman
writes what the feature must do
3
bsman
challenges the story before any code
4
designman
writes the design you approve
5
tddman
writes the tests, then the code
6
docman
updates the docs to what shipped
for example, an epic that compares todo-list algorithms:
  1. epicman create an epic for comparing todo-list algorithms, and its storieswrites the epic and five stories, benchmark among them
  2. bsman on benchmarkasks the open questions, each with a proposition and its rationale
  3. storyman updatefolds your answers into the story
  4. designman on benchmarkplans the code from what the repository holds today
  5. tddman on benchmarkwrites the tests, sees them fail, then writes the code
  6. docman on benchmarkcompletes the documentation
  7. storyman submitcloses the story: it no longer blocks the others
license PolyForm Noncommercial 1.0.0: free for noncommercial use.
open code: github.com/robusta-build/rb-compound
// open questions

every step ends on questions.

an agent never guesses what you did not say. when the epic, a story, a brainstorm or a design leaves a choice open, its agent writes an open question at the end of the file: the question, the one answer it proposes, and why. you answer in a line, and the next agent starts from your answer. one step after the other, the project gets sharper.

OQ-BENCH-2 how many items does the benchmark measure?

Proposition:
1,000, 10,000 and 100,000 items.
Rationale:
three orders of magnitude show where an algorithm stops scaling, and a run stays under a minute.
Resolution:
lgtm

OQ-BENCH-3 which algorithms go into the first comparison?

Proposition:
an array, a linked list and a balanced tree.
Rationale:
they cover the three costs of an insertion: shifting, linking, rebalancing.
Resolution:
keep the array and the list, add a skip list instead of the tree.

lgtm, “looks good to me”: the proposition is taken as it is. any other answer is yours, in your own words.

// boss, the complete version

one command, the whole pipeline.

boss runs the six agents in your place, one after the other, story after story. it answers the open questions their propositions settle, signs each answer, and leaves you the ones that need you. it stops where you tell it to, and logs every decision in one file.

boss
1
epicman
the epic
2
storyman
the story
3
bsman
the questions
4
designman
the design
5
tddman
the tests, the code
6
docman
the docs
the same epic, with boss:
  1. epicman create an epic for comparing todo-list algorithms, and its storieswrites the epic and five stories
  2. boss atw todo-list-algorithmsruns the five stories to the commit, answers the open questions and logs why
this version only comes with a paid, personalized accompaniment of your team.
// the rules of the workflow

rules, written down.

let’s talk about your team.

Tell me how your team builds software with agents today. I will tell you whether the accompaniment fits.

No form, no newsletter: I read and answer every message myself.